Two nights ago a colleague pinged me on Slack. He’s digging into content marketing and asked a simple question: what AI tools are you using for content?
I started typing an answer. Ten minutes later I was still typing — connectors, folder systems, moodboards, skills, the image model that keeps a face consistent, the video platform that gives you six models under one roof. Somewhere around the fourth message I wrote, “honestly, I could talk about this for hours,” and realized that’s what a field guide is for.
So here it is: the whole stack, with the specifics. Not the tools I’ve heard are good — the ones that were open on my machine this morning at the coffee shop in Fernandina Beach.
The short answer first
If you only read one paragraph: Claude Code is the workhorse, and everything else is a hand it can reach. It sets up the projects, drafts the copy, builds the brand guides, calls the image model, calls the video platform, ships the site. The tools that matter most in my stack aren’t generators at all — they’re the context I feed the workhorse: project folders, voice guidelines, brand guidelines, interview transcripts. And at every step there’s a human — me — approving, redirecting, killing the weak frames. The stack doesn’t make the content good. The stack makes the loop fast enough that taste gets to run more laps.
Now the specifics.
Rule one: context beats tools
Ask me what makes AI content actually work and the answer isn’t a tool. It’s context — feeding the right material to whatever tool you’re using, and having a system so that material is always in reach.
Concretely, that means a tight project folder system. One folder per client or brand. Inside each: a short operating file that says what this project is and how it works, a voice guide (how this person or brand actually sounds — built from their real writing and interviews, not adjectives), a brand guide (palette, type, the rules that are ratified and the ones still open), and the working files. When I open Claude Code inside that folder, it starts already knowing the world it’s working in. The output is pointed at the idea instead of generic — and generic is the failure mode of every AI content pipeline.
The real system, real shape — client names anonymized, structure verbatim. One folder per client or brand; the operating file, voice guide, and brand guide live at the top of each.
I wrote a whole guide on this idea — The Company Is the Context — so I won’t re-argue it here. The one-line version: the labs that build these models all converged on the same lesson, which is that managing what the model sees matters more than the prompt. Your folder system is how a one-person shop does that.
The workhorse: Claude Code
Claude Code is a command-line app — you talk to it in plain English and it works inside your actual files. That description undersells it. In practice it’s the layer everything else routes through:
- Setting up projects. New client, new campaign, new series — it scaffolds the folder, drafts the operating file, and the structure is real from day one.
- Building voice and brand guidelines, fast. This used to be a weeks-long deliverable. Now it’s: point it at everything the client has written, plus the interviews, and shape the guide together in a sitting. The speed isn’t the point — the point is the guide exists, so every piece after it starts from voice instead of from zero.
- Leveraging interviews. This is the underrated one. Interviews are the raw ore of good content — the stories, the phrases people actually use, the thing they said at minute forty-one that becomes the headline. I record calls with tl;dv or Granola, and Claude Code pulls the transcripts straight in through connectors. An hour of conversation becomes a voice guide, a post series, a case study — without the transcript ever passing through my clipboard.
An hour of conversation becomes the standard — interview in through the connector, voice guide out, a human reading it against the real thing before it becomes law.
The design wing: Claude Design, moodboards, and skills
For building websites, my ship-it layer is Netlify — but the part people don’t expect is what comes before the code: moodboarding has been the secret.
Before I build anything visual, I gather references — screenshots, sites, textures — and react to them. What I like, what I don’t, in plain words. That reacting is the design brief. It turns “make it feel premium” (useless) into named ingredients a machine can actually build from. Claude Design has been spectacular at taking that board and turning it into real pages — but only because the board did the taste work first.
The other unlock: skills — small instruction packs that teach Claude a discipline so you’re not re-explaining your standards every session. Three I lean on for design work: a taste skill that bans the templated AI-slop look, one that encodes Emil Kowalski’s philosophy of UI polish and the invisible details, and one called Impeccable that audits and refines an interface like a senior design reviewer. Load them and the floor of what comes out rises immediately.
The honest miss: Figma. Everyone asks. I haven’t found the Claude Code ↔ Figma handoff to work well for this kind of content-and-copy work — the loop of moodboard → Claude Design → real page on Netlify is faster for me than routing through design files. Maybe that changes; for now I go around it.
The picture-and-motion wing: Higgsfield + Nano Banana
For images and video, the combination that changed things for me is Higgsfield connected with Nano Banana.
Nano Banana (Google’s image model) is the art department — its superpower is consistency, keeping the same face, place, or object intact while everything around it changes. Higgsfield is the motion studio — and the quiet advantage is that it’s a multi-model platform: it gives you access to Seedance and a whole roster of other engines under one roof, so you’re not betting your pipeline on a single model’s taste. Its own image and content generation is fantastic too.
Both connect to Claude Code, which means the whole thing runs as one conversation: brief → stills → my approval → motion → finished cuts in every aspect ratio. I wrote the full setup guide here: How to Set Up Higgsfield with Claude Code.
Off this exact line — three frames from the Van Gogh series (“what if Van Gogh painted us — and our machines?”), made with Nano Banana driven from Claude Code, curated down from seventeen to nine by the taste gate.
I know the fashionable answer for animation right now is Remotion — video as React code. I’ll be honest: for content work, I don’t think that’s where the leverage is. If you’re a content person, the still-to-motion pipeline above will get you to finished, publishable video with taste gates the whole way. Code-built animation is a different craft for a different job.
What holds it together isn’t in the stack
Here’s the thing the tool list hides. Every pipeline above has the same shape: the machine proposes, the human decides. I pick the frames before anything animates. I react to the moodboard before anything gets built. I read the voice guide against the person’s real writing before it becomes law.
That’s the actual answer to my colleague’s question. The tools are genuinely great — and they change; half this list didn’t exist eighteen months ago. What doesn’t change: context in, taste on top. Build the folder system. Write the voice guides. Hold the gate. The stack just makes the loop fast enough for that to be your whole job — which, if you make content for a living, it always was.
— Joe
